A research team has built an AI system that generates synthetic PET brain scans from regular MRI data, a cheaper and radiation-free stand-in for a test that is central to dementia diagnosis.
The framework, called DB-SUiT, works directly on the folded surface of the cortex rather than treating the brain as a generic 3D volume, since that is where dementia-related metabolic changes actually show up. It pairs a spherical convolutional encoder with a transformer to model both local surface detail and long-range patterns across the brain, and it factors in a patient's demographics and subcortical structure to refine the output. Tested on two datasets covering different dementia types, the synthetic scans boosted automated diagnostic accuracy by 14.2% over using MRI alone and 11.3% over standard volume-based PET synthesis. In a blinded reading test, human readers hit 85.5% diagnostic accuracy on the synthetic scans, versus 75.8% for MRI and 95.2% for real PET.
FDG-PET is one of the most reliable dementia biomarkers available, but it is expensive, involves radiation exposure, and is out of reach for many clinics, especially outside major hospital systems. A tool that approximates it from MRI, which is far more widely available, could extend earlier dementia screening to patients who would otherwise never get a PET scan. The model also held up on an external cohort that included a dementia subtype it had never seen during training, a meaningful test of whether this generalizes beyond a curated dataset.
Still, a 10-point accuracy gap between synthetic and real PET is not nothing when the diagnosis is dementia, and this is a research paper with code on GitHub, not a cleared clinical tool.